1. What are the main areas of research in machine learning applications for SAT solving?
The main areas of research in machine learning applications for SAT solving are creating standalone SAT solvers with pure machine learning methods, replacing components of existing conflict-driven clause learning (CDCL) solvers with learning-directed heuristics, and modifying local search solvers with learning-aided modules. These areas focus on improving the efficiency and accuracy of SAT solving techniques by integrating machine learning algorithms. Standalone SAT solvers with pure machine learning methods aim to develop SAT solvers that solely rely on machine learning techniques for solving SAT problems. Replacing components of existing CDCL solvers with learning-directed heuristics involves incorporating machine learning algorithms to guide the solver's decision-making process, enhancing its ability to find solutions more efficiently. Modifying local search solvers with learning-aided modules involves integrating machine learning algorithms to assist in the local search process, improving the solver's ability to explore the solution space effectively. These research areas contribute to advancing the field of SAT solving by leveraging the power of machine learning to enhance problem-solving capabilities.
read more
2. How is the random forest trained?
The random forest is trained on hand-crafted instances from the CBS dataset from SATlib. The dataset comprises 1000 random-3-SAT CNFs with controlled backbone size, each with 100 variables and 403 clauses. 80 instances are probed to generate a labelled dataset with a variety of sizes and balanced distribution between SAT and UNSAT. The branching can be directed by the solutions computed by the SAT solver to obtain a labelled dataset with SAT/UNSAT balance. A random forest classifier is then hyperparametrised and trained on this dataset. Finally, the classifier is deployed in step 5. of the iterative approach. The model is evaluated using statistical SATzilla features (ML) and a model that includes probing features (ML+probing).
read more
3. How does the Davis-Putnam Pure Literal Rule contribute to variable selection?
The Davis-Putnam Pure Literal Rule focuses on purity in variable selection. It aims to select a variable that, if assigned incorrectly, makes the resulting formula unsatisfiable. This rule considers the binary characteristic of strict purity, but also introduces a measure to evaluate how close a literal is to purity and its importance for formula resolution. By selecting the variable with the highest score, calculated using the formula score(v) = |f (l) - f (!l)|, where l is the positive literal and !l is the negated literal, the rule ensures that the chosen variable has a significant impact on the formula's satisfiability. The function f, which measures the literal relevance and purity, includes the number of appearances of the literal, its covariance with the negations of crucial literals, and the average clause size in which the literal appears. This comprehensive approach helps in identifying the most critical variable for resolving the formula effectively.
read more
4. How does Majority heuristic compare to Random assignment?
The Majority heuristic significantly outperforms the Random assignment in the Results section. It assigns variables to completion 28% of the test instances, while the Random assignment performs poorly, making an invalid assignment early. The Majority heuristic uses a simple approach of fixing the variable to false if the literal mostly appears negated. This heuristic strongly outperforms the classifier based on statistical features. Further investigations on alternative variable selection heuristics, classifiers, and extracted features are suggested to improve performance. The generalization capabilities should be assessed by increasing the size and variety of the test set. The pipeline could also be improved, including the implementation of backtracking.
read more